Přepis řeči - On the use of GSV-SVM for Speaker Diarization and Tracking

0:00:06	so
0:00:07	hi everybody and
0:00:08	but the high
0:00:10	i i'm presenting the world
0:00:11	on the use of jesse as yen for speaker diarization and tracking
0:00:16	that was brought on by my colleague yeah exactly
0:00:19	great income and i come to the conference
0:00:21	income collaboration with not too hard
0:00:24	i can see generous
0:00:25	before mac please to uh
0:00:28	we live in
0:00:28	happen
0:00:31	so after the presentation of the of the task and the with division i will describe the two tasks
0:00:36	uh
0:00:38	that are explored
0:00:39	i quickly speaker diarization and speaker tracking
0:00:42	along with this
0:00:42	stance
0:00:43	uh
0:00:44	uh use than the result obtained before compression and plastic
0:00:50	first about the the two tasks that we consider a acoustic speaker diarisation is a
0:00:55	who who spoke when' task
0:00:57	we got married segmentation and clustering that was really and right
0:01:00	the previews
0:01:01	um
0:01:02	uh
0:01:03	oh
0:01:04	um
0:01:05	we consider it as a
0:01:06	processing for the uh
0:01:09	automatic speech recognition in for underage
0:01:12	transcription
0:01:13	uh in this situation we have new approach
0:01:15	priori information on speaker phone
0:01:17	and speaker's voice
0:01:19	and we can see there you only acoustic driven approaches
0:01:22	because al also approach
0:01:24	linguistic use of the of the transcription
0:01:26	but we are also interested in just to get writing that
0:01:30	uh
0:01:31	well we want to detect regional for spoken documents that
0:01:34	a detailed by a given speaker
0:01:36	and this situation we have a list of the speaker to the right
0:01:40	and we have uh
0:01:41	we are provided with training that that for
0:01:44	for this because
0:01:45	we consider the speaker tracking task as a combination of both acoustic speaker diarization
0:01:51	press
0:01:51	the speaker verification module in our configuration
0:01:56	oh i'm and motivation in this work was to uh
0:02:00	including the our system the
0:02:02	as yeah techniques that
0:02:04	you know all that they are
0:02:06	become
0:02:07	very uh successfully in
0:02:09	speaker recognition
0:02:11	uh
0:02:12	we started with the
0:02:13	G S B
0:02:15	questions about vectors
0:02:16	stan
0:02:17	that was
0:02:17	easy to develop a framework
0:02:19	and they also for the
0:02:21	uh
0:02:22	well the features that can be used
0:02:24	uh in uh
0:02:25	uh
0:02:26	as the end system for speaker recognition that well
0:02:29	mllr cmllr lattice mllr that
0:02:32	uh we also uh
0:02:34	uh
0:02:35	efficient to combine with the the
0:02:37	do not you as it is
0:02:39	or more
0:02:40	i want also to say one about the context of the we're programs that
0:02:44	uh
0:02:45	what do you rate
0:02:45	our work for improving
0:02:48	uh
0:02:48	so it's a a friend found that the
0:02:51	research
0:02:52	uh
0:02:52	and uh
0:02:53	uh
0:02:54	innovation program
0:02:56	uh
0:02:56	that
0:02:57	aims to improve automatic would you be sure uh document structuring and indexing
0:03:01	and
0:03:02	for this work we wanted to work specifically on the speaker diarization
0:03:06	and tracking for
0:03:07	exactly what 'cause that yeah
0:03:09	and that's why we are uh still
0:03:12	on the us team of
0:03:14	uh
0:03:14	offline
0:03:15	um
0:03:16	diarisation
0:03:17	uh because we we are working on a carry on broadcast that that that are recorded and that
0:03:23	uh patch on the on the web or on the radio on T V
0:03:26	so one
0:03:27	and we will
0:03:29	still easy integration of
0:03:30	been based
0:03:31	fig
0:03:34	we worked on the
0:03:35	that out of the
0:03:36	french ester evaluation
0:03:39	uh
0:03:40	i hope that the that that will
0:03:42	so softly available to the
0:03:44	were community
0:03:45	remedied as being distributed to the participant to this evaluation in two thousand eight
0:03:50	uh yeah uh
0:03:51	one hundred
0:03:52	being a target speakers
0:03:54	uh for which uh
0:03:56	we have
0:03:57	about
0:03:57	one and the right
0:03:58	i well as a
0:03:59	training that that
0:04:02	consist of
0:04:03	french speaking or radio shows
0:04:05	from
0:04:06	uh
0:04:08	different sources
0:04:09	french tools but also uh
0:04:11	uh
0:04:12	i consoles
0:04:14	uh
0:04:15	we have
0:04:17	for the impostor that that we've to the
0:04:19	it's yeah one which is uh
0:04:21	uh
0:04:22	is that when evaluation uh that that
0:04:24	about
0:04:24	four hundred
0:04:25	impostors
0:04:27	uh
0:04:28	as the two development data consisting twenty radio show for that a lot
0:04:32	six hours and the evaluation
0:04:34	uh a row roughly the same amount
0:04:37	twenty six radio shows
0:04:38	uh false
0:04:39	seven hours
0:04:41	uh i also provide uh
0:04:43	uh
0:04:44	and the value of the
0:04:46	uh
0:04:46	so if you use some statistics
0:04:48	on the number of speakers as speaking anything
0:04:51	the segment length
0:04:52	the
0:04:53	uh S yeah
0:04:54	uh
0:04:54	to development and evaluation uh that that's that
0:04:58	uh
0:04:58	the development we we have between nine and twenty five
0:05:02	speaker for the mean of C
0:05:04	being
0:05:04	and
0:05:05	on the
0:05:06	evaluation
0:05:07	uh
0:05:08	roughly lies
0:05:10	uh with the
0:05:10	given uh
0:05:13	speaker there so we
0:05:15	in a right
0:05:16	the speaking length
0:05:17	also vary a lot
0:05:19	with a mean of uh
0:05:21	sixty five seconds
0:05:23	ranging between alpha signal and then
0:05:25	more than ten minutes
0:05:27	and on the evaluation the
0:05:29	it's a bit
0:05:29	i yeah with
0:05:30	and the right of
0:05:31	it is signals
0:05:32	but we can see the standard deviation is very very high so it
0:05:36	just
0:05:36	just to have a rough
0:05:38	uh i
0:05:39	and segments
0:05:40	oh also
0:05:41	yeah
0:05:42	in average
0:05:43	six
0:05:43	yeah
0:05:44	seventeen seconds and
0:05:46	the deadline for you know the about it
0:05:48	we also uh rummaging from a fraction of signal to
0:05:52	uh so norman
0:05:56	i will not describe the uh acoustic speaker diarization system
0:06:00	uh which is basically this is ten that uh high guy
0:06:03	uh recap recap just
0:06:05	the previous stork
0:06:06	that was developed by uh
0:06:08	C than men yeah uh
0:06:10	changed to uh myself and only woman
0:06:13	for them is
0:06:14	two thousand for evaluation
0:06:16	uh
0:06:17	basically uh well
0:06:20	so we just that
0:06:21	uh
0:06:22	die revisions
0:06:23	then
0:06:24	so initial a segmentation is using a front end with
0:06:28	standard mfcc feature found in uh is a system
0:06:31	uh
0:06:32	the speech activity detection
0:06:35	relies on the viterbi decoding with
0:06:37	that's ever uh gmms of
0:06:39	speech music and noise
0:06:43	on the speech segment
0:06:44	there is
0:06:45	um
0:06:46	uh
0:06:48	the segmentation to ins more uh segments using
0:06:53	to select
0:06:53	two i just some sliding windows of output signal
0:06:57	and using a local gardens major to segment
0:07:01	i does that uh
0:07:03	gmm are trained on the signal
0:07:06	yeah the lower segmentation
0:07:08	uh of this uh of this data
0:07:11	this is the initial segmentation
0:07:13	we have the first step of
0:07:15	uh i
0:07:15	i dramatic clustering using
0:07:18	using the
0:07:19	classical bic italian
0:07:21	uh
0:07:22	and
0:07:22	oh using full covariance matrix on the single version
0:07:27	uh something only
0:07:29	uh specific
0:07:30	thing is on the
0:07:31	penalty which is the local
0:07:33	big in I T
0:07:34	uh taking into account only the number
0:07:37	um it out of the
0:07:39	two clusters that are um
0:07:41	but
0:07:41	and not of the
0:07:42	all that
0:07:43	and not
0:07:44	or of the future
0:07:48	and we uh
0:07:50	with
0:07:51	put the output of the biclustering into a second step
0:07:54	using
0:07:56	uh
0:07:56	speaker I D um
0:07:58	mode that is and uh clustering
0:08:00	using
0:08:02	a slightly different
0:08:05	features
0:08:06	using
0:08:07	uh feature warping
0:08:09	and not adapting the ubm
0:08:12	the clustering relies
0:08:14	the force log likelihood ratio between the two clusters
0:08:21	so what what we did the
0:08:22	uh was
0:08:23	but there are simple
0:08:24	simple stuff was looking at the G as the as yet
0:08:28	then
0:08:28	and
0:08:29	integrate it into the system
0:08:31	place of the uh last
0:08:33	S I Ds clustering stage
0:08:36	uh so i think i
0:08:39	whiskey
0:08:39	right up
0:08:40	first of all that it's
0:08:42	rose asked on that
0:08:43	stuff
0:08:44	in
0:08:45	the G S V U
0:08:48	consist of the means of the
0:08:50	uh
0:08:50	so predictor of the adapted gmm
0:08:57	it
0:08:58	exactly as that
0:08:59	uh combining diarisation system can improve on the
0:09:04	individual assistance
0:09:06	there are several ways of doing this combination
0:09:08	uh
0:09:09	i think one system into the ozone
0:09:11	the kind of thing that we already do in our
0:09:14	stan
0:09:14	we can also mounts different systems or
0:09:17	do a
0:09:17	cluster voting technique
0:09:19	we did a version that's
0:09:21	score label
0:09:23	which means that
0:09:24	during the
0:09:25	clustering
0:09:27	process
0:09:29	we compute an average score between the G and then
0:09:32	as the end
0:09:33	and the
0:09:33	G and then
0:09:34	geodesy and then gmmubm schools
0:09:38	with
0:09:38	uh
0:09:39	the weight
0:09:39	oh that optimise
0:09:40	on the that
0:09:42	development
0:09:45	the performance measure is the diarization error rate
0:09:49	it was already described in uh
0:09:51	preview stored so i want
0:09:53	get
0:09:53	too much into
0:09:54	then
0:09:55	uh again
0:09:56	uh just
0:09:57	to say that we also
0:09:59	put some
0:10:00	i could use with
0:10:01	she with your coverage
0:10:02	which are the ratio of the minutes reference speaker about within this cluster
0:10:07	and
0:10:08	combats ripples of possible right
0:10:10	which can provide a
0:10:11	better insight into the
0:10:13	the speaker or
0:10:15	and we use the
0:10:16	the nist to for scoring following the
0:10:19	step two evaluation plan
0:10:23	yeah
0:10:24	yeah he's a figure of the diarisation error right
0:10:28	four
0:10:28	uh
0:10:29	i was
0:10:30	ten
0:10:30	uh
0:10:31	to the to the left
0:10:32	we have the performance of the gmm
0:10:34	then
0:10:35	to the right of the the G S E S P N
0:10:37	then
0:10:38	and
0:10:39	uh
0:10:40	on the uh
0:10:41	X axis
0:10:42	the different combination weight
0:10:45	we have been
0:10:46	right
0:10:47	in the
0:10:48	in green
0:10:49	the green curve is for the evaluation set and the right go
0:10:53	for the development set
0:10:55	which is that the
0:10:56	gmm is yeah
0:10:58	yeah forms
0:10:59	better than the
0:11:01	that that the gmmubm so we have a bit of that
0:11:03	the that yeah this is then
0:11:05	but combination uh
0:11:07	yeah
0:11:07	is very uh
0:11:09	successfully here
0:11:11	uh
0:11:12	more in detail
0:11:13	what what we get is
0:11:14	a ten percent
0:11:16	relative improvements
0:11:17	from a given that one to ten not one
0:11:20	from the
0:11:21	best performing
0:11:22	then to the
0:11:23	the G and then
0:11:25	press
0:11:25	at the end
0:11:26	then
0:11:27	uh on the development set
0:11:29	and
0:11:29	on the evaluation we are also saying
0:11:31	right
0:11:32	going down from uh nine that
0:11:34	six two
0:11:35	it but
0:11:36	three
0:11:42	was this was for the acoustic speaker diarization system
0:11:45	no some words about the the speaker tracking
0:11:49	as i said we uh you just stand as a combination of
0:11:53	uh of the speaker acoustic speaker diarization system to the to the left
0:11:57	with a speaker at our educations
0:11:59	stan
0:12:00	and we have
0:12:02	three possible ways of doing this
0:12:04	combination
0:12:05	uh we can do this you can verification
0:12:08	on the initial segments
0:12:10	of the system
0:12:11	or on the cluster
0:12:13	output by the beach
0:12:15	all by the S I D
0:12:16	uh clustering step
0:12:19	each case
0:12:20	the segments are then
0:12:21	uh compared to the
0:12:22	speaker models and
0:12:23	level according
0:12:30	if you well
0:12:31	a weekly on the on the C stands for the tracking system
0:12:35	we use
0:12:35	gmmubm and she is this the end system
0:12:38	uh is that a
0:12:40	uh
0:12:42	that have the same uh
0:12:44	properties as as assistant that we
0:12:46	that we have already presented folder
0:12:49	i musician uh a
0:12:51	uh for the verification we choose the target model
0:12:54	with the highest likelihood
0:12:56	ratio
0:12:58	i with the verification phase
0:13:00	and
0:13:00	the G S via the end is also uh
0:13:03	is also
0:13:04	following the the same
0:13:08	the same uh
0:13:10	architecture
0:13:12	uh
0:13:13	with the constraint that we scrolls input
0:13:16	posters and target
0:13:17	because
0:13:18	uh using the agenda
0:13:20	and channel matching the the current condition
0:13:25	and we also perform
0:13:26	uh waited at the right of but
0:13:28	across all the
0:13:29	so level
0:13:29	uh system fusion
0:13:33	the the performance measures for the
0:13:34	tracking task where as
0:13:36	finding the exact way evaluation campaign
0:13:39	uh
0:13:39	recall and precision
0:13:41	and
0:13:42	an issue of combining the
0:13:44	but recall and precision
0:13:46	um
0:13:49	but in a time waiting
0:13:51	um i manner
0:13:52	and also the speaker weighted action that was the proposed
0:13:56	doing the the F and you're that speaker
0:14:03	we have something on the on the debt curls that that was
0:14:06	simulated by uh by using a
0:14:09	uh
0:14:10	short segments of the evaluation data
0:14:13	of all the different possible uh
0:14:15	so
0:14:16	then
0:14:17	on the evaluation on that then and on the
0:14:20	yeah
0:14:21	so you're on the evaluation data
0:14:25	but suppose that you and then you yeah and the G S V as yeah
0:14:28	then
0:14:29	we shall with the red the green and the blue gel
0:14:33	a different version of the gmmubm since then
0:14:36	uh with the
0:14:37	verification applied
0:14:39	i sat at the output of the segment
0:14:42	initial segmentation in blue
0:14:44	as the output of the
0:14:45	D
0:14:45	stayed in a red
0:14:46	of the output of the excited state in green
0:14:50	it appears that there are not that much if you're ounces and we add
0:14:53	a slightly better
0:14:55	a four months
0:14:56	by using the output of the final stage
0:14:58	which is a exciting stage
0:15:00	and the G is yes yes
0:15:02	then
0:15:03	uh
0:15:03	yeah you don't shown on the on this
0:15:05	uh i'll put on the S I D the clustering step
0:15:08	and uh
0:15:10	yeah
0:15:10	is
0:15:10	you shown to perform much better than the
0:15:13	gmmubm
0:15:17	yeah some some figures i i will consult
0:15:19	uh provides a recall precision if you're right
0:15:22	uh
0:15:23	and
0:15:24	if your average by by speaker
0:15:26	i i will mainly call uh focus on the uh
0:15:30	S
0:15:31	uh problem of the of the result
0:15:34	that that of him on the
0:15:36	uh on the dev and on T of or on the about it
0:15:39	well what we
0:15:41	what we
0:15:42	so on the on the go
0:15:44	on the
0:15:45	on the development set
0:15:46	we
0:15:47	uh
0:15:48	observe that the
0:15:49	S I D clustering step
0:15:51	provide the
0:15:53	uh
0:15:54	better performance that that was a condition
0:15:56	and at the
0:15:57	compare able uh performance
0:15:59	that of the G S the end
0:16:01	then
0:16:02	and the combination uh is uh
0:16:04	improving upon but it's then
0:16:07	on the user on on on the evaluation dataset
0:16:10	uh the G S V as yeah and you
0:16:13	performing much better than the G bit gmmubm
0:16:16	then
0:16:17	and this case
0:16:19	the combination
0:16:20	a slightly outperform the G S T I E N
0:16:23	and is
0:16:24	better than the gmm ubm
0:16:26	then
0:16:29	well
0:16:29	uh
0:16:30	that was uh
0:16:32	i would say a a simple
0:16:33	experimental framework
0:16:35	uh that
0:16:37	was
0:16:38	done that to do the integration of the D N A is yeah
0:16:41	then into uh
0:16:42	speaker diarization and tracking
0:16:44	then
0:16:45	so just yeah yeah as
0:16:46	a private school
0:16:48	performance to the
0:16:49	existing standard gmm ubm
0:16:52	uh that we that we had
0:16:54	and the
0:16:55	the score level fusion was uh
0:16:58	what's that
0:16:59	factory
0:17:00	uh there are uh some caveats
0:17:02	yeah
0:17:03	for example in the post all set
0:17:05	which is not very balanced
0:17:06	according to the
0:17:07	gender and the channel
0:17:09	that
0:17:10	there are some very small set for example
0:17:13	honestly made on our bound that that we have a very few posters for the
0:17:17	uh for the experiments
0:17:19	and of course
0:17:20	we want to go
0:17:22	browser for
0:17:23	well the svm features
0:17:24	like an L L F
0:17:25	cmllr lattice mllr
0:17:27	and also the very interesting direction that were presented in the
0:17:31	previous bill
0:17:34	thank you for that
0:17:44	cool
0:17:47	yeah and that is that
0:17:49	you using delta double delta features
0:17:52	yeah
0:17:53	uh and
0:17:54	some other posters
0:17:56	found that the amending the deltas
0:17:58	yeah
0:18:00	it is
0:18:01	you could addendum limiting the deltas
0:18:03	so
0:18:03	the
0:18:04	did you
0:18:05	you try limiting the deltas
0:18:07	yeah
0:18:08	for example in the
0:18:09	in the first
0:18:10	stayed on the beach
0:18:12	segmentation
0:18:13	uh
0:18:15	on the initial segmentation we use the delta delta
0:18:18	on the first
0:18:19	uh that uh
0:18:20	on the big stage we use of
0:18:22	comments metrics using
0:18:23	nothing at all or only the
0:18:25	a static features
0:18:28	and on the second stage we use
0:18:30	uh only the delta
0:18:32	not the delta delta
0:18:34	um
0:18:35	the the russian one of the rationale is
0:18:38	trying to have different
0:18:39	uh
0:18:40	feature representation
0:18:41	to combine different
0:18:43	uh aspects
0:18:44	a different flavour
0:18:45	i i'm not sure that it is optimal this way
0:18:47	because we wouldn't test or configuration
0:18:49	it was one way of doing that
0:18:51	but i i i agree that that that that that that
0:18:55	it's not
0:18:56	clearly convincing that they bring uh always something uh in the district
0:19:11	i
0:19:11	to to to what people say that we observe that when the data very clean
0:19:16	yeah
0:19:17	we make some recruiting
0:19:19	acoustic room
0:19:20	so the didn't uh didn't they give some game but i mean
0:19:24	it's
0:19:24	data from nineteen
0:19:26	precision you will never see any given to the different user
0:19:34	thank you
0:19:51	you don't
0:19:51	slide fourteen years
0:19:53	forty
0:19:53	slide forty
0:19:55	okay yeah
0:19:58	yes you can explain why you could input result and
0:20:01	the evaluation did there was a difference between a database
0:20:06	the databases are
0:20:07	where is that well not recorded at the same day
0:20:10	and
0:20:11	they have a slightly different balance between the sources
0:20:14	of data
0:20:15	there are some some
0:20:16	uh
0:20:17	so that coming from french
0:20:19	i'll send their
0:20:20	some of the uh from uh
0:20:22	i pupils and that's good not
0:20:24	was uh from a high view mall
0:20:25	was up so they are
0:20:27	and the balancing is different between the that when the evil
0:20:30	slightly different and i think it's
0:20:33	it's
0:20:33	uh and it's
0:20:34	blind some
0:20:36	some
0:20:36	reside
0:20:37	and also the fact that uh
0:20:39	well
0:20:41	i think that for the acquisition system even the weight is uh
0:20:46	the givens images
0:20:47	even twenty uh
0:20:49	well
0:20:50	six a well it's not that much
0:20:54	and then
0:20:55	when you do speaker tracking
0:20:57	you have like sometimes
0:21:00	speaker
0:21:00	we speak a lot
0:21:01	so his model is what is fine
0:21:03	sometimes speaker lies only few
0:21:05	time to
0:21:05	however on the speaker
0:21:07	so how to fix it yourself in this case
0:21:11	um
0:21:17	on the speaker tracking it's a verification
0:21:20	uh
0:21:22	compared to
0:21:24	two of racial
0:21:25	so there is only a
0:21:27	the normalisation by by the length but the the the free should there is no
0:21:31	normalisation uh according to the length of the of the data
0:21:34	i i agree that it is some something that needs to be uh
0:21:38	addressed
0:21:39	yeah
0:21:51	and this
0:21:51	i think yeah not

On the use of GSV-SVM for Speaker Diarization and Tracking

SESSION 6: Diarization

Přidáno: 14. 7. 2010 11:08, Autor: Viet Bac Le (LIMSI-CNRS), Claude Barras (LIMSI-CNRS, Univ. Paris-Sud), Marc Ferras (LIMSI-CNRS), Délka: 0:21:58